cd /news/ai-agents/loopback-is-not-a-trust-boundary-for… · home › topics › ai-agents › article
[ARTICLE · art-173035] src=dev.to ↗ pub= topic=ai-agents verified=true sentiment=↓ negative

Loopback Is Not a Trust Boundary for Todoist's HTTP MCP Server

Todoist MCP versions through 10.3.2 shipped an optional HTTP server that accepted browser-origin MCP requests without validating Host or Origin headers or requiring a caller secret, allowing DNS rebinding to route a malicious public site to the victim's loopback listener and execute tools with the configured TODOIST_API_KEY. The advisory, tracked as CVE-2026-76807, demonstrated task creation and completion and notes a wider write surface including updates, deletion, projects, comments, reminders and assignments; the stdio entrypoint was not affected. Users are urged to upgrade @doist/todoist-mcp to 10.3.3 or later, prefer stdio for local agent use, and rotate API keys if the HTTP server was reachable by an untrusted browser.

recommendation

Validate the caller, not just the socket address: enforce explicit allowed hosts and origins plus a per-run token on any local MCP HTTP server, and prefer stdio for local agent use.

more in playbook: Securing agents and AI tools
by read1 min views2 publishedNov 20, 2026

Local AI tool servers are often treated as safe because they bind to 127.0.0.1. CVE-2026-76807 shows why network location and caller identity are different controls.

Todoist MCP versions through 10.3.2 exposed an optional todoist-mcp-http server. The server accepted browser-origin MCP requests without validating Host or Origin and without requiring a caller secret. DNS rebinding allowed a malicious public website to route its own hostname to the victim's loopback listener.

The server then executed tools with its configured TODOIST_API_KEY. The advisory demonstrated task creation and completion, and describes the wider write surface: updates, deletion, projects, sections, comments, reminders, filters, labels, and assignments. The stdio entrypoint is not browser-reachable and was not affected by this attack path.

Upgrade @doist/todoist-mcp to 10.3.3 or later. Stop unnecessary HTTP transport instances and prefer stdio for local agent use. If HTTP mode is required, enforce strict loopback hostnames, explicit allowed origins, and a per-run caller token.

Rebinding protections in an MCP SDK are useful only when the server passes explicit allowed-host and allowed-origin lists. Do not assume that loopback binding enables them automatically.

Search Todoist activity for tasks, comments, projects, or deletions created through automation without a matching user action. Review logs for MCP requests with unexpected Host or Origin values and for browser sessions that resolved to loopback.

Rotate the Todoist API key if the HTTP server was reachable by an untrusted browser or if anomalous writes occurred. Revoke local bearer tokens that were exposed.

Treat every local tool server as an internet-adjacent service when a browser can issue requests to it. Validate the caller, not just the socket address. Keep high-impact write tools behind explicit authorization and preserve an audit trail that names the tool, caller session, and resulting change.

── more in #ai-agents 4 stories · sorted by recency
── more on @todoist 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/loopback-is-not-a-tr…] indexed:0 read:1min 2026-11-20 · —